01 / 17
Week 4

Context, memory & usage

The recording layer habit, the week that separates people who try Codex for a fortnight from people who build on it.

Works.
Concept

The context window is one box. Input and output share it.

System prompt, everything you've typed, everything Codex has read, plus everything it writes back, all stacking up in the same space.

System prompt User Assistant User Assistant Input solid, output dashed, one shared box
Concept

Hit the limit and the message cuts off, sometimes mid sentence.

Codex cannot see past its own wall, and it will not warn you it's coming.

Message 1 Message 2 Message 3 Hard limit Message 4, cut off
Concept, named phenomenon

The start and the end get attention. The middle fades.

Fewer tokens, chosen well, beat more tokens dumped in.

Start Middle, fades End
Concept, the spine

Retrieval is the operating layer. Capture is a storage unit.

Give every task the smallest context that completes it, a token budget, not a firehose.

Feedback Permissions Source truth Retrieval Capture Most teams stop at Capture and call it a brain
Concept

Context cascades top down: company, team, project, individual.

At Happy Haus that cascade runs through the SharePoint folder structure, know which level you're editing before you assume it flows down.

Concept

Anyone can demo the first run. We get judged on the second.

If the second run isn't faster, better, or usable by someone other than the person who built it, nothing was installed.

Run 1 Hovering Explaining Correcting Run 2 Unassisted, faster
Worked example

Kat's weekly construction minutes are the two minute job that never happens.

Record, transcribe, draft, Kat reviews and sends.

Site supervisor track

Site walkthroughs are where the supervisor track plugs in.

Bruce and Jai's voice notes on site become drafted daily log entries and task updates.

Governance

The recording habit starts with consent, not a demo.

One page consent and staff notification, signed off with Rob before this week opens. We read from Jack and Databuild once the read-only exports land. We never write into either.

Worked example

Rob's Weekly Ops Brief is the strategic layer Jack doesn't produce today.

Assembled by hand or not at all, until the recording habit and the read-only export are both live.

Concept

Debrief writes what happened. Reflect writes what you learned.

Reflect runs before debrief, every session, while the correction is still fresh. Both write to the SharePoint-hosted instructions file for that unit, demoed on Codex.

Concept

A knowledge base nobody prunes is a junk drawer with a search bar.

Stale context is worse than no context, because Codex believes it.

Concept

Tag every number hard or soft, and never let a soft number anchor a headline claim.

Frame any recovered-time figure as a working estimate to be confirmed, never a promise, and set the precedence per workflow in writing.

Concept

The rest of the brain fills the same way every time: parse, classify, extract, canonicalise, author, validate.

Smoke test five to ten high signal items first, get sign-off, then broaden, never all in one sitting.

Before next session

This week's work

  • Kat: record this week's construction meeting minutes, let Codex transcribe and draft them, then review and send them herself.
  • Bruce and Jai: capture one week of voice notes from site walkthroughs as the baseline sample for daily logs and task updates.
  • Rob: sign off the one page consent and staff notification before anyone hits record.
  • Everyone in the core cohort: install Debrief and Reflect, and run reflect before debrief at the end of every working session this week.
Exit criteria

You're done when

  • Reflect ran before debrief in every session this week.
  • No usage limit was hit, checked against Codex's current usage windows rather than assumed.
  • Kat's weekly minutes were recorded, transcribed, drafted, and sent by her at least once.
  • Bruce and Jai's baseline week of voice notes is captured, even if it's unmeasured until now.
  • One stale file or entry was pruned, and someone can say what Codex would have believed if it hadn't been.